The document presents a framework combining accelerated particle swarm optimization (APSO) and support vector machines (SVM) for addressing business optimization challenges. The authors demonstrate the framework's utility through case studies in production optimization, income prediction, and project scheduling. The paper highlights the efficiency of APSO and SVM in processing large datasets for enhancing business performance under constraints.

1. Accelerated Particle Swarm Optimization and
Support Vector Machine for Business
Optimization and Applications
Xin-She Yang1
, Suash Deb2
, and Simon Fong3
1
Department of Engineering, University of Cambridge, Trumpinton Street,
Cambridge CB2 1PZ, UK
xy227@cam.ac.uk
2
Department of Computer Science & Engineering, C.V. Raman College of
Engineering, Bidyanagar, Mahura, Janla, Bhubaneswar 752054, India
suashdeb@gmail.com
3
Department of Computer and Information Science, Faculty of Science and
Technology, University of Macau, Taipa, Macau
ccfong@umac.mo
Abstract. Business optimization is becoming increasingly important
because all business activities aim to maximize the profit and perfor-
mance of products and services, under limited resources and appropriate
constraints. Recent developments in support vector machine and meta-
heuristics show many advantages of these techniques. In particular, par-
ticle swarm optimization is now widely used in solving tough optimiza-
tion problems. In this paper, we use a combination of a recently devel-
oped Accelerated PSO and a nonlinear support vector machine to form
a framework for solving business optimization problems. We first apply
the proposed APSO-SVM to production optimization, and then use it for
income prediction and project scheduling. We also carry out some para-
metric studies and discuss the advantages of the proposed metaheuristic
SVM.
Keywords: Accelerated PSO, business optimization, metaheuristics, PSO,
support vector machine, project scheduling.
1 Introduction
Many business activities often have to deal with large, complex databases. This
is partly driven by information technology, especially the Internet, and partly
driven by the need to extract meaningful knowledge by data mining. To ex-
tract useful information among a huge amount of data requires efficient tools
for processing vast data sets. This is equivalent to trying to find an optimal so-
lution to a highly nonlinear problem with multiple, complex constraints, which
is a challenging task. Various techniques for such data mining and optimization
S. Fong et al. (Eds.): NDT 2011, CCIS 136, pp. 53–66, 2011.
c Springer-Verlag Berlin Heidelberg 2011

2. 54 X.-S. Yang, S. Deb, and S. Fong
have been developed over the past few decades. Among these techniques, support
vector machine is one of the best techniques for regression, classification and data
mining [5,9,16,19,20,24].
On the other hand, metaheuristic algorithms also become powerful for solv-
ing tough nonlinear optimization problems [1,7,8,27,32]. Modern metaheuristic
algorithms have been developed with an aim to carry out global search, typi-
cal examples are genetic algorithms [6], particle swarm optimisation (PSO) [7],
and Cuckoo Search [29,30]. The efficiency of metaheuristic algorithms can be
attributed to the fact that they imitate the best features in nature, especially
the selection of the fittest in biological systems which have evolved by natural
selection over millions of years. Since most data have noise or associated random-
ness, most these algorithms cannot be used directly. In this case, some form of
averaging or reformulation of the problem often helps. Even so, most algorithms
become difficult to implement for such type of optimization.
In addition to the above challenges, business optimization often concerns with
a large amount but often incomplete data, evolving dynamically over time. Cer-
tain tasks cannot start before other required tasks are completed, such complex
scheduling is often NP-hard and no universally efficient tool exists. Recent trends
indicate that metaheuristics can be very promising, in combination with other
tools such as neural networks and support vector machines [5,9,15,21].
In this paper, we intend to present a simple framework of business optimiza-
tion using a combination of support vector machine with accelerated PSO. The
paper is organized as follows: We first will briefly review particle swarm opti-
mization and accelerated PSO, and then introduce the basics of support vector
machines (SVM). We then use three case studies to test the proposed frame-
work. Finally, we discussion its implications and possible extension for further
research.
2 Accelerated Particle Swarm Optimization
2.1 PSO
Particle swarm optimization (PSO) was developed by Kennedy and Eberhart
in 1995 [7,8], based on the swarm behaviour such as fish and bird schooling in
nature. Since then, PSO has generated much wider interests, and forms an ex-
citing, ever-expanding research subject, called swarm intelligence. PSO has been
applied to almost every area in optimization, computational intelligence, and
design/scheduling applications. There are at least two dozens of PSO variants,
and hybrid algorithms by combining PSO with other existing algorithms are also
increasingly popular.
PSO searches the space of an objective function by adjusting the trajectories of
individual agents, called particles, as the piecewise paths formed by positional
vectors in a quasi-stochastic manner. The movement of a swarming particle
consists of two major components: a stochastic component and a deterministic
component. Each particle is attracted toward the position of the current global

3. PSO and SVM for Business Optimization 55
best g∗
and its own best location x∗
i in history, while at the same time it has a
tendency to move randomly.
Let xi and vi be the position vector and velocity for particle i, respectively.
The new velocity vector is determined by the following formula
vt+1
i = vt
i + α 1[g∗
− xt
i] + β 2[x∗
i − xt
i]. (1)
where 1 and 2 are two random vectors, and each entry taking the values between
0 and 1. The parameters α and β are the learning parameters or acceleration
constants, which can typically be taken as, say, α ≈ β ≈ 2.
There are many variants which extend the standard PSO algorithm, and the
most noticeable improvement is probably to use an inertia function θ(t) so that
vt
i is replaced by θ(t)vt
i
vt+1
i = θvt
i + α 1[g∗
− xt
i] + β 2[x∗
i − xt
i], (2)
where θ ∈ (0, 1) [2,3]. In the simplest case, the inertia function can be taken as
a constant, typically θ ≈ 0.5 ∼ 0.9. This is equivalent to introducing a virtual
mass to stabilize the motion of the particles, and thus the algorithm is expected
to converge more quickly.
2.2 Accelerated PSO
The standard particle swarm optimization uses both the current global best g∗
and the individual best x∗
i . The reason of using the individual best is primarily
to increase the diversity in the quality solutions, however, this diversity can be
simulated using some randomness. Subsequently, there is no compelling reason
for using the individual best, unless the optimization problem of interest is highly
nonlinear and multimodal.
A simplified version which could accelerate the convergence of the algorithm is
to use the global best only. Thus, in the accelerated particle swarm optimization
(APSO) [27,32], the velocity vector is generated by a simpler formula
vt+1
i = vt
i + α n + β(g∗
− xt
i), (3)
where n is drawn from N(0, 1) to replace the second term. The update of the
position is simply
xt+1
i = xt
i + vt+1
i . (4)
In order to increase the convergence even further, we can also write the update
of the location in a single step
xt+1
i = (1 − β)xt
i + βg∗
+ α n. (5)
This simpler version will give the same order of convergence. Typically, α =
0.1L ∼ 0.5L where L is the scale of each variable, while β = 0.1 ∼ 0.7 is
sufficient for most applications. It is worth pointing out that velocity does not
appear in equation (5), and there is no need to deal with initialization of velocity

4. 56 X.-S. Yang, S. Deb, and S. Fong
vectors. Therefore, APSO is much simpler. Comparing with many PSO variants,
APSO uses only two parameters, and the mechanism is simple to understand.
A further improvement to the accelerated PSO is to reduce the randomness
as iterations proceed. This means that we can use a monotonically decreasing
function such as
α = α0e−γt
, (6)
or
α = α0γt
, (0 < γ < 1), (7)
where α0 ≈ 0.5 ∼ 1 is the initial value of the randomness parameter. Here t is
the number of iterations or time steps. 0 < γ < 1 is a control parameter [32].
For example, in our implementation, we will use
α = 0.7t
, (8)
where t ∈ [0, tmax] and tmax is the maximum of iterations.
3 Support Vector Machine
Support vector machine (SVM) is an efficient tool for data mining and classifica-
tion [25,26]. Due to the vast volumes of data in business, especially e-commerce,
efficient use of data mining techniques becomes a necessity. In fact, SVM can
also be considered as an optimization tool, as its objective is to maximize the
separation margins between data sets. The proper combination of SVM with
metaheuristics could be advantageous.
3.1 Support Vector Machine
A support vector machine essentially transforms a set of data into a significantly
higher-dimensional space by nonlinear transformations so that regression and
data fitting can be carried out in this high-dimensional space. This methodology
can be used for data classification, pattern recognition, and regression, and its
theory was based on statistical machine learning theory [21,24,25].
For classifications with the learning examples or data (xi, yi) where i =
1, 2, ..., n and yi ∈ {−1, +1}, the aim of the learning is to find a function φα(x)
from allowable functions {φα : α ∈ Ω} such that φα(xi) → yi for (i = 1, 2, ..., n)
and that the expected risk E(α) is minimal. That is the minimization of the risk
E(α) =
1
2
|φα(x) − y|dQ(x, y), (9)
where Q(x, y) is an unknown probability distribution, which makes it impossible
to calculate E(α) directly. A simple approach is to use the so-called empirical
risk
Ep(α) ≈
1
2n
n
i=1
φα(xi) − yi . (10)

5. PSO and SVM for Business Optimization 57
However, the main flaw of this approach is that a small risk or error on the
training set does not necessarily guarantee a small error on prediction if the
number n of training data is small [26].
For a given probability of at least 1 − p, the Vapnik bound for the errors can
be written as
E(α) ≤ Rp(α) + Ψ
h
n
,
log(p)
n
, (11)
where
Ψ
h
n
,
log(p)
n
=
1
n
h(log
2n
h
+ 1) − log(
p
4
) . (12)
Here h is a parameter, often referred to as the Vapnik-Chervonenskis dimension
or simply VC-dimension [24], which describes the capacity for prediction of the
function set φα.
In essence, a linear support vector machine is to construct two hyperplanes
as far away as possible and no samples should be between these two planes.
Mathematically, this is equivalent to two equations
w · x + b = ±1, (13)
and a main objective of constructing these two hyperplanes is to maximize the
distance (between the two planes)
d =
2
||w||
. (14)
Such maximization of d is equivalent to the minimization of ||w|| or more conve-
niently ||w||2
. From the optimization point of view, the maximization of margins
can be written as
minimize
1
2
||w||2
=
1
2
(w · w). (15)
This essentially becomes an optimization problem
minimize Ψ =
1
2
||w||2
+ λ
n
i=1
ηi, (16)
subject to yi(w · xi + b) ≥ 1 − ηi, (17)
ηi ≥ 0, (i = 1, 2, ..., n), (18)
where λ > 0 is a parameter to be chosen appropriately. Here, the term
n
i=1 ηi
is essentially a measure of the upper bound of the number of misclassifications
on the training data.
3.2 Nonlinear SVM and Kernel Tricks
The so-called kernel trick is an important technique, transforming data dimen-
sions while simplifying computation. By using Lagrange multipliers αi ≥ 0, we

6. 58 X.-S. Yang, S. Deb, and S. Fong
can rewrite the above constrained optimization into an unconstrained version,
and we have
L =
1
2
||w||2
+ λ
n
i=1
ηi −
n
i=1
αi[yi(w · xi + b) − (1 − ηi)]. (19)
From this, we can write the Karush-Kuhn-Tucker conditions
∂L
∂w
= w −
n
i=1
αiyixi = 0, (20)
∂L
∂b
= −
n
i=1
αiyi = 0, (21)
yi(w · xi + b) − (1 − ηi) ≥ 0, (22)
αi[yi(w · xi + b) − (1 − ηi)] = 0, (i = 1, 2, ..., n), (23)
αi ≥ 0, ηi ≥ 0, (i = 1, 2, ..., n). (24)
From the first KKT condition, we get
w =
n
i=1
yiαixi. (25)
It is worth pointing out here that only the nonzero αi contribute to overall
solution. This comes from the KKT condition (23), which implies that when
αi = 0, the inequality (17) must be satisfied exactly, while α0 = 0 means the
inequality is automatically met. In this latter case, ηi = 0. Therefore, only the
corresponding training data (xi, yi) with αi > 0 can contribute to the solution,
and thus such xi form the support vectors (hence, the name support vector
machine). All the other data with αi = 0 become irrelevant.
It has been shown that the solution for αi can be found by solving the following
quadratic programming [24,26]
maximize
n
i=1
αi −
1
2
n
i,j=1
αiαjyiyj(xi · xj), (26)
subject to
n
i=1
αiyi = 0, 0 ≤ αi ≤ λ, (i = 1, 2, ..., n). (27)
From the coefficients αi, we can write the final classification or decision function
as
f(x) = sgn
n
i=1
αiyi(x · xi) + b , (28)
where sgn is the classic sign function.

7. PSO and SVM for Business Optimization 59
As most problems are nonlinear in business applications, and the above linear
SVM cannot be used. Ideally, we should find some nonlinear transformation
φ so that the data can be mapped onto a high-dimensional space where the
classification becomes linear. The transformation should be chosen in a certain
way so that their dot product leads to a kernel-style function K(x, xi) = φ(x) ·
φ(xi). In fact, we do not need to know such transformations, we can directly use
the kernel functions K(x, xi) to complete this task. This is the so-called kernel
function trick. Now the main task is to chose a suitable kernel function for a
given, specific problem.
For most problems in nonlinear support vector machines, we can use
K(x, xi) = (x·xi)d
for polynomial classifiers, K(x, xi) = tanh[k(x·xi)+Θ)] for
neural networks, and by far the most widely used kernel is the Gaussian radial
basis function (RBF)
K(x, xi) = exp −
||x − xi||2
(2σ2)
= exp − γ||x − xi||2
, (29)
for the nonlinear classifiers. This kernel can easily be extended to any high
dimensions. Here σ2
is the variance and γ = 1/2σ2
is a constant. In general, a
simple bound of 0 < γ ≤ C is used, and here C is a constant.
Following the similar procedure as discussed earlier for linear SVM, we can
obtain the coefficients αi by solving the following optimization problem
maximize
n
i=1
αi −
1
2
αiαjyiyjK(xi, xj). (30)
It is worth pointing out under Mercer’s conditions for the kernel function, the
matrix A = yiyjK(xi, xj) is a symmetric positive definite matrix [26], which
implies that the above maximization is a quadratic programming problem, and
can thus be solved efficiently by standard QP techniques [21].
4 Metaheuristic Support Vector Machine with APSO
4.1 Metaheuristics
There are many metaheuristic algorithms for optimization and most these al-
gorithms are inspired by nature [27]. Metaheuristic algorithms such as genetic
algorithms and simulated annealing are widely used, almost routinely, in many
applications, while relatively new algorithms such as particle swarm optimiza-
tion [7], firefly algorithm and cuckoo search are becoming more and more pop-
ular [27,32]. Hybridization of these algorithms with existing algorithms are also
emerging.
The advantage of such a combination is to use a balanced tradeoff between
global search which is often slow and a fast local search. Such a balance is im-
portant, as highlighted by the analysis by Blum and Roli [1]. Another advantage
of this method is that we can use any algorithms we like at different stages of

8. 60 X.-S. Yang, S. Deb, and S. Fong
the search or even at different stage of iterations. This makes it easy to combine
the advantages of various algorithms so as to produce better results.
Others have attempted to carry out parameter optimization associated with
neural networks and SVM. For example, Liu et al. have used SVM optimized
by PSO for tax forecasting [13]. Lu et al. proposed a model for finding optimal
parameters in SVM by PSO optimization [14]. However, here we intend to pro-
pose a generic framework for combining efficient APSO with SVM, which can
be extended to other algorithms such as firefly algorithm [28,31].
4.2 APSO-SVM
Support vector machine has a major advantage, that is, it is less likely to overfit,
compared with other methods such as regression and neural networks. In addi-
tion, efficient quadratic programming can be used for training support vector
machines. However, when there is noise in the data, such algorithms are not
quite suitable. In this case, the learning or training to estimate the parameters
in the SVM becomes difficult or inefficient.
Another issue is that the choice of the values of kernel parameters C and σ2
in the kernel functions; however, there is no agreed guideline on how to choose
them, though the choice of their values should make the SVM as efficiently as
possible. This itself is essentially an optimization problem.
Taking this idea further, we first use an educated guess set of values and use
the metaheuristic algorithms such as accelerated PSO or cuckoo search to find
the best kernel parameters such as C and σ2
[27,29]. Then, we used these param-
eters to construct the support vector machines which are then used for solving
the problem of interest. During the iterations and optimization, we can also mod-
ify kernel parameters and evolve the SVM accordingly. This framework can be
called a metaheuristic support vector machine. Schematically, this Accelerated
PSO-SVM can be represented as shown in Fig. 1.
begin
Define the objective;
Choose kernel functions;
Initialize various parameters;
while (criterion)
Find optimal kernel parameters by APSO;
Construct the support vector machine;
Search for the optimal solution by APSO-SVM;
Increase the iteration counter;
end
Post-processing the results;
end
Fig. 1. Metaheuristic APSO-SVM.

9. PSO and SVM for Business Optimization 61
For the optimization of parameters and business applications discussed below,
APSO is used for both local and global search [27,32].
5 Business Optimization Benchmarks
Using the framework discussed earlier, we can easily implement it in any pro-
gramming language, though we have implemented using Matlab. We have vali-
dated our implementation using the standard test functions, which confirms the
correctness of the implementation. Now we apply it to carry out case studies with
known analytical solution or the known optimal solutions. The Cobb-Douglas
production optimization has an analytical solution which can be used for com-
parison, while the second case is a standard benchmark in resource-constrained
project scheduling [11].
5.1 Production Optimization
Let us first use the proposed approach to study the classical Cobb-Douglas pro-
duction optimization. For a production of a series of products and the labour
costs, the utility function can be written
q =
n
j=1
u
αj
j = uα1
1 uα2
2 · · · uαn
n , (31)
where all exponents αj are non-negative, satisfying
n
j=1
αj = 1. (32)
The optimization is the minimization of the utility
minimize q (33)
subject to
n
j=1
wjuj = K, (34)
where wj(j = 1, 2, ..., n) are known weights.
This problem can be solved using the Lagrange multiplier method as an un-
constrained problem
ψ =
n
j=1
u
αj
j + λ(
n
j=1
wjuj − K), (35)
whose optimality conditions are
∂ψ
∂uj
= αju−1
j
n
j=1
u
αj
j + λwj = 0, (j = 1, 2, ..., n), (36)

10. 62 X.-S. Yang, S. Deb, and S. Fong
∂ψ
∂λ
=
n
j=1
wjuj − K = 0. (37)
The solutions are
u1 =
K
w1[1 + 1
α1
n
j=2 αj]
, uj =
w1αj
wjα1
u1, (38)
where (j = 2, 3, ..., n). For example, in a special case of n = 2, α1 = 2/3,
α2 = 1/3, w1 = 5, w2 = 2 and K = 300, we have
u1 =
Q
w1(1 + α2/α1)
= 40, u2 =
Kα2
w2α1(1 + α2/α1)
= 50.
As most real-world problem has some uncertainty, we can now add some noise
to the above problem. For simplicity, we just modify the constraint as
n
j=1
wjuj = K(1 + β ), (39)
where is a random number drawn from a Gaussian distribution with a zero
mean and a unity variance, and 0 ≤ β 1 is a small positive number.
We now solve this problem as an optimization problem by the proposed APSO-
SVM. In the case of β = 0.01, the results have been summarized in Table 1 where
the values are provided with different problem size n with different numbers of
iterations. We can see that the results converge at the optimal solution very
quickly.
Table 1. Mean deviations from the optimal solutions
size n Iterations deviations
10 1000 0.014
20 5000 0.037
50 5000 0.040
50 15000 0.009
6 Income Prediction
Studies to improve the accuracy of classifications are extensive. For example,
Kohavi proposed a decision-tree hybrid in 1996 [10]. Furthermore, an efficient
training algorithm for support vector machines was proposed by Platt in 1998
[17,18], and it has some significant impact on machine learning, regression and
data mining.
A well-known benchmark for classification and regression is the income predic-
tion using the data sets from a selected 14 attributes of a household from a sensus

11. PSO and SVM for Business Optimization 63
form [10,17]. We use the same data sets at ftp://ftp.ics.uci.edu/pub/machine-
learning-databases/adult for this case study. There are 32561 samples in the
training set with 16281 for testing. The aim is to predict if an individual’s in-
come is above or below 50K ?
Among the 14 attributes, a subset can be selected, and a subset such as age,
education level, occupation, gender and working hours are commonly used.
Using the proposed APSO-SVM and choosing the limit value of C as 1.25, the
best error of 17.23% is obtained (see Table 2), which is comparable with most
accurate predictions reported in [10,17].
Table 2. Income prediction using APSO-SVM
Train set (size) Prediction set Errors (%)
512 256 24.9
1024 256 20.4
16400 8200 17.23
6.1 Project Scheduling
Scheduling is an important class of discrete optimization with a wider range of
applications in business intelligence. For resource-constrained project scheduling
problems, there exists a standard benchmark library by Kolisch and Sprecher
[11,12]. The basic model consists of J activities/tasks, and some activities cannot
start before all its predecessors h are completed. In addition, each activity j =
1, 2, ..., J can be carried out, without interruption, in one of the Mj modes, and
performing any activity j in any chosen mode m takes djm periods, which is
supported by a set of renewable resource R and non-renewable resources N.
The project’s makespan or upper bound is T, and the overall capacity of non-
renewable resources is Kν
r where r ∈ N. For an activity j scheduled in mode
m, it uses kρ
jmr units of renewable resources and kν
jmr units of non-renewable
resources in period t = 1, 2, ..., T.
For activity j, the shortest duration is fit into the time windows [EFj, LFj]
where EFj is the earliest finish times, and LFj is the latest finish times. Math-
ematically, this model can be written as [11]
Minimize Ψ(x)
Mj
m=1
LFj
t=EFj
t · xjmt, (40)
subject to
Mh
m=1
LFj
t=EFj
t xhmt ≤
Mj
m=1
LFj
t=EFj
(t − djm)xjmt, (j = 2, ..., J),
J
j=1
Mj
m=1
kρ
jmr
min{t+djm−1,LFj}
q=max{t,EFj }
xjmq ≤ Kρ
r , (r ∈ R),

12. 64 X.-S. Yang, S. Deb, and S. Fong
J
j=1
Mj
m=1
kν
jmr
LFj
t=EFj
xjmt ≤ Kν
r , (r ∈ N), (41)
and
Mj
j=1
t = EFj
LFj
= 1, j = 1, 2, ..., J, (42)
where xjmt ∈ {0, 1} and t = 1, ..., T. As xjmt only takes two values 0 or 1,
this problem can be considered as a classification problem, and metaheuristic
support vector machine can be applied naturally.
Table 3. Kernel parameters used in SVM
Number of iterations SVM kernel parameters
1000 C = 149.2, σ2
= 67.9
5000 C = 127.9, σ2
= 64.0
Using the online benchmark library [12], we have solved this type of problem
with J = 30 activities (the standard test set j30). The run time on a modern
desktop computer is about 2.2 seconds for N = 1000 iterations to 15.4 seconds
for N = 5000 iterations. We have run the simulations for 50 times so as to obtain
meaningful statistics.
The optimal kernel parameters found for the support vector machines are
listed in Table 3, while the deviations from the known best solution are given in
Table 4 where the results by other methods are also compared.
Table 4. Mean deviations from the optimal solution (J=30)
Algorithm Authors N = 1000 5000
PSO [22] Kemmoe et al. (2007) 0.26 0.21
hybribd GA [23] Valls eta al. (2007) 0.27 0.06
Tabu search [15] Nonobe & Ibaraki (2002) 0.46 0.16
Adapting GA [4] Hartmann (2002) 0.38 0.22
Meta APSO-SVM this paper 0.19 0.025
From these tables, we can see that the proposed metaheuristic support vector
machine starts very well, and results are comparable with those by other meth-
ods such as hybrid genetic algorithm. In addition, it converges more quickly,
as the number of iterations increases. With the same amount of function eval-
uations involved, much better results are obtained, which implies that APSO
is very efficient, and subsequently the APSO-SVM is also efficient in this con-
text. In addition, this also suggests that this proposed framework is appropriate
for automatically choosing the right parameters for SVM and solving nonlinear
optimization problems.

13. PSO and SVM for Business Optimization 65
7 Conclusions
Both PSO and support vector machines are now widely used as optimization
techniques in business intelligence. They can also be used for data mining to
extract useful information efficiently. SVM can also be considered as an opti-
mization technique in many applications including business optimization. When
there is noise in data, some averaging or reformulation may lead to better per-
formance. In addition, metaheuristic algorithms can be used to find the optimal
kernel parameters for a support vector machine and also to search for the op-
timal solutions. We have used three very different case studies to demonstrate
such a metaheuristic SVM framework works.
Automatic parameter tuning and efficiency improvement will be an important
topic for further research. It can be expected that this framework can be used
for other applications. Furthermore, APSO can also be used to combine with
other algorithms such as neutral networks to produce more efficient algorithms
[13,14]. More studies in this area are highly needed.
References
1. Blum, C., Roli, A.: Metaheuristics in combinatorial optimization: Overview and
conceptural comparision. ACM Comput. Surv. 35, 268–308 (2003)
2. Chatterjee, A., Siarry, P.: Nonlinear inertia variation for dynamic adaptation in
particle swarm optimization. Comp. Oper. Research 33, 859–871 (2006)
3. Clerc, M., Kennedy, J.: The particle swarm - explosion, stability, and convergence
in a multidimensional complex space. IEEE Trans. Evolutionary Computation 6,
58–73 (2002)
4. Hartmann, S.: A self-adapting genetic algorithm for project scheduling under
resource constraints. Naval Res. Log. 49, 433–448 (2002)
5. Howley, T., Madden, M.G.: The genetic kernel support vector machine: descrip-
tion and evaluation. Artificial Intelligence Review 24, 379–395 (2005)
6. Goldberg, D.E.: Genetic Algorithms in Search, Optimisation and Machine Learn-
ing. Addison Wesley, Reading (1989)
7. Kennedy, J., Eberhart, R.C.: Particle swarm optimization, in. In: Proc. of IEEE
International Conference on Neural Networks, Piscataway, NJ, pp. 1942–1948
(1995)
8. Kennedy, J., Eberhart, R.C.: Swarm intelligence. Academic Press, London (2001)
9. Kim, K.: Financial forecasting using support vector machines. Neurocomput-
ing 55, 307–319 (2003)
10. Kohavi, R.: Scaling up the accuracy of naive-Bayes classifiers: a
decision-tree hybrid. In: Proc. 2nd Int. Conf. on Knowledge Discov-
ery and Data Mining, pp. 202–207. AAAI Press, Menlo Park (1996),
ftp://ftp.ics.uci.edu/pub/machine-learning-databases/adult
11. Kolisch, R., Sprecher, A.: PSPLIB - a project scdeluing problem library, OR
Software-ORSEP (operations research software exchange prorgam) by H. W.
Hamacher. Euro. J. Oper. Res. 96, 205–216 (1996)
12. Kolisch, R., Sprecher, A.: The Library PSBLIB,
http://129.187.106.231/psplib/

14. 66 X.-S. Yang, S. Deb, and S. Fong
13. Liu, L.-X., Zhuang, Y., Liu, X.Y.: Tax forecasting theory and model based on
SVM optimized by PSO. Expert Systems with Applications 38, 116–120 (2011)
14. Lu, N., Zhou, J.Z., He, Y.: Y., Liu Y., Particle Swarm Optimization for Parameter
Optimization of Support Vector Machine Model. In: 2009 Second International
Conference on Intelligent Computation Technology and Automation, pp. 283–284.
IEEE publications, Los Alamitos (2009)
15. Nonobe, K., Ibaraki, T.: Formulation and tabu search algorithm for the resource
constrained project scheduling problem (RCPSP). In: Ribeiro, C.C., Hansen, P.
(eds.) Essays and Surveys in Metaheuristics, pp. 557–588 (2002)
16. Pai, P.F., Hong, W.C.: Forecasting regional electricity load based on recurrent
support vector machines with genetic algorithms. Electric Power Sys. Res. 74,
417–425 (2005)
17. Platt, J.C.: Sequential minimal optimization: a fast algorithm for training support
vector machines, Techical report MSR-TR-98014, Microsoft Research (1998)
18. Plate, J.C.: Fast training of support vector machines using sequential minimal
optimization, in. In: Scholkopf, B., Burges, C.J., Smola, A.J. (eds.) Advances in
Kernel Methods – Support Vector Learning, pp. 185–208. MIT Press, Cambridge
(1999)
19. Shi, G.R.: The use of support vector machine for oil and gas identification in
low-porosity and low-permeability reservoirs. Int. J. Mathematical Modelling and
Numerical Optimisation 1, 75–87 (2009)
20. Shi, G.R., Yang, X.-S.: Optimization and data mining for fracture prediction in
geosciences. Procedia Computer Science 1, 1353–1360 (2010)
21. Smola, A. J., Sch¨olkopf, B.: A tutorial on support vector regression, (1998),
http://www.svms.org/regression/
22. Tchomt´e, S.K., Gourgand, M., Quilliot, A.: Solving resource-constrained project
scheduling problem with particle swarm optimization. In: Proceeding of 3rd Mul-
tidsciplinary Int. Scheduling Conference (MISTA 2007), Paris, August 28 - 31,
pp. 251–258 (2007)
23. Valls, V., Ballestin, F., Quintanilla, S.: A hybrid genetic algorithm for the
resource-constrained project scheduling problem. Euro. J. Oper. Res (2007),
doi:10.1016/j.ejor.2006.12.033
24. Vapnik, V.: Estimation of Dependences Based on Empirical Data. Springer, New
York (1982) (in Russian)
25. Vapnik, V.: The nature of Statistical Learning Theory. Springer, New York (1995)
26. Scholkopf, B., Sung, K., Burges, C., Girosi, F., Niyogi, P., Poggio, T., Vapnik, V.:
Comparing support vector machine with Gaussian kernels to radial basis function
classifiers. IEEE Trans. Signal Processing 45, 2758–2765 (1997)
27. Yang, X.S.: Nature-Inspired Metaheuristic Algorithms. Luniver Press (2008)
28. Yang, X.-S.: Firefly algorithms for multimodal optimization. In: Watanabe, O.,
Zeugmann, T. (eds.) SAGA 2009. LNCS, vol. 5792, pp. 169–178. Springer, Hei-
delberg (2009)
29. Yang, X.-S., Deb, S.: Cuckoo search via L´evy flights, in. In: Proceeings of World
Congress on Nature & Biologically Inspired Computing, NaBIC 2009, pp. 210–
214. IEEE Publications, USA (2009)
30. Yang, X.S., Deb, S.: Engineering optimization by cuckoo search. Int. J. Mathe-
matical Modelling and Numerical Optimisation 1, 330–343 (2010)
31. Yang, X.S.: Firefly algorithm, stochastic test functions and design optimisation.
Int. J. Bio-inspired Computation 2, 78–84 (2010)
32. Yang, X.S.: Engineering Optimization: An Introduction with Metaheuristic Ap-
plications. John Wiley & Sons, Chichester (2010)